MoGloss: 2D keypoints, captions, pseudo-glosses, splits and model weights

This repository holds the data and the trained models of MoGloss, a motion encoder for American Sign Language (ASL) pretrained with masked motion modeling and with joint motion and pseudo-gloss modeling. It contains

• 2D whole-body keypoints for about 1.1 million continuous-signing clips (YouTube-ASL, YouTube-SL-25 ASL subset, OpenASL, ASL STEM Wiki, How2Sign) and for the isolated-sign datasets ASL Citizen, ASL SignBank and WLASL2000,
• captions and pseudo-glosses for every continuous clip,
• the train / validation / test lists of every experiment in the paper,
• the model weights behind every result of the paper, for our model and for our reproductions of Uni-Sign and the YouTube-ASL baseline.

Paper: arXiv:XXXX.XXXXX (link will be updated). Code: https://github.com/hongrui16/MoGloss

If you use the keypoints, captions, pseudo-glosses, splits or weights of this repository, please cite our paper (see Citation).

Layout

keypoints/
  rtmpose_2d/<dataset>/<dataset>-NNNNN.tar     RTMPose keypoints, one <clip>.npz per clip
  mediapipe/mediapipe-NNNNN.tar                MediaPipe landmarks, one <clip>.npz per clip
metadata/<dataset>.jsonl                        one line per clip (captions, pseudo-glosses, labels)
splits/
  setting_a/*.txt  setting_b/*.txt             translation experiments
  openasl_test_seen.txt  openasl_test_unseen.txt
  islr/*.txt                                    isolated sign recognition
weights/
  mogloss/{setting_a,setting_b,islr}/...       our models
  reproductions/{uni-sign,youtube_asl_baseline}/...
  MANIFEST.tsv                                  path, sha256 and size of every weight file

Keypoints

RTMPose (keypoints/rtmpose_2d)

All clips were processed with one pose estimator, the lightweight whole-body configuration of rtmlib (YOLOX-tiny person detector and RTMW-m at 256x192), on every second video frame. Each <clip>.npz holds

key	shape	content
keypoints	(T, 133, 2)	x / frame width, y / frame height (the Uni-Sign pose format), COCO-WholeBody order
scores	(T, 133)	keypoint confidence

Our model and all keypoint-based reproductions read these files. The paper uses 69 of the 133 points (9 body, 21 per hand, 18 face, the selection of Uni-Sign) with the Uni-Sign normalization.

folder	clips	content
youtube_asl	232,129	YouTube-ASL caption segments after sentence merging and filtering (setting B)
youtube_asl_raw_captions	603,243	YouTube-ASL caption cues as released (setting A)
youtube_sl25_ase	83,181	ASL videos of YouTube-SL-25 that are not in YouTube-ASL
openasl	94,097	OpenASL sentences
openasl_setting_a_box	6,617	multi-person OpenASL clips re-extracted inside the official signer box, used in setting A (see below)
how2sign	35,126	How2Sign sentence clips, front view
asl_stem_wiki	63,552	ASL STEM Wiki sentences
asl_citizen	83,381	ASL Citizen isolated signs (5 training videos without keypoints are missing)
asl_signbank	2,836	ASL SignBank citation-form videos with keypoints
wlasl2000	21,083	WLASL2000 files as released
wlasl2000_annotation_cut	21,082	the same files cut to the annotated span (input of our Uni-Sign WLASL reproduction)

OpenASL in setting A. For 6,617 multi-person OpenASL clips, setting A reads the box-guided extraction in openasl_setting_a_box, and setting B reads the plain extraction in openasl. Both folders use the same clip names.

MediaPipe (keypoints/mediapipe)

Input of our SHuBERT reproduction: MediaPipe pose, hand and face landmarks, following the SHuBERT preprocessing. 792,991 clips (the clips of setting A). Each <clip>.npz holds pose (T, 25, 2), lhand (T, 21, 2), rhand (T, 21, 2), face (T, 13, 2), the signer box bbox (4,), the frame size hw (2,), the clip length T and per-part detection counts (mp_pose, mp_lhand, mp_rhand, mp_face, mp_res_fail).

Clip names

Clip names are our internal ids with : replaced by __. The prefix gives the dataset.

dataset	example
YouTube-ASL (setting B)	yt__<video>_0000 (video id and segment number)
YouTube-ASL raw captions	ytseg__<video>__<start>_<end> (seconds)
YouTube-SL-25	sl25__<video>_0000
OpenASL	oa__<video>-HH__MM__SS.mmm-HH__MM__SS.mmm (OpenASL timestamps with : replaced by __)
How2Sign	h2s__<official clip name>
ASL STEM Wiki	stem__<id>
ASL Citizen / SignBank / WLASL	original file names

The MediaPipe files use the same names, so the two keypoint sets can be joined by name.

Metadata

metadata/<dataset>.jsonl, one JSON object per clip.

Continuous datasets:

{"id": "oa__0HzztPiFelY-00__11__15.673-00__11__16.018", "corpus": "openasl", "video_id": "0HzztPiFelY", "n_frames": 46,
 "setting_a": {"caption": "...", "pseudo_gloss": null},
 "setting_b": {"caption": "...", "pseudo_gloss": ["...", "..."]}}

setting_a / setting_b are present when the clip belongs to that setting. caption is the English sentence used for translation. pseudo_gloss is the list of pseudo-gloss units used in joint pretraining (null for clips that are not in the pretraining data). Pseudo-glosses were produced by Qwen3.5-27B from the captions, with a lexicon of ASL SignBank and ASL Citizen glosses in the prompt.

Isolated datasets: gloss, and for ASL Citizen signer and split, for WLASL2000 signer, split and in_wlasl100. ASL Citizen and WLASL2000 also give islr_span = [start, end), the frames our classifiers read (the annotated span for uncut WLASL files, then the frames in which a hand is visible, widened by 2 frames).

Splits

Setting A (translation pretraining on YouTube-ASL, then fine-tuning per dataset): pretrain_youtube_asl_train (602,243), pretrain_youtube_asl_dev (1,000) and its 600-clip checkpoint-selection subset, finetune_{how2sign,stem,openasl}_train, test_{how2sign,stem,openasl}.

Both settings also list the clips of joint pretraining (joint_train, joint_val) and of translation pretraining (translation_train, after dropping clips shorter than 0.64 seconds); setting B adds translation_train_{50,25}pct.

Setting B (one model on all data, no fine-tuning): train (472,246), train_50pct, train_25pct, val (2,000) and val_600, test_{how2sign,stem,openasl,youtube_asl}. Every clip linked to a test video of any dataset is removed from the setting-B training data.

OpenASL seen / unseen. 608 of the 908 OpenASL test sentences come from videos that are also in YouTube-ASL. openasl_test_seen.txt and openasl_test_unseen.txt split the test set accordingly.

ISLR: islr/asl_citizen_{train,val,test} (official split, test = 32,928 clips of 11 unseen signers) and islr/wlasl{100,2000}_{train,val,test} (official WLASL v0.3 split). Clip 70053 of WLASL has no frames and is skipped by the Uni-Sign recipe.

Reading the data

import io, json, tarfile, numpy as np

with tarfile.open("keypoints/rtmpose_2d/openasl/openasl-00000.tar") as tf:
    for m in tf:
        clip = m.name[:-4]                         # strip ".npz"
        z = np.load(io.BytesIO(tf.extractfile(m).read()))
        kp, sc = z["keypoints"], z["scores"]       # (T, 133, 2), (T, 133)
        break

meta = {r["id"]: r for r in map(json.loads, open("metadata/openasl.jsonl"))}
test = open("splits/setting_b/test_openasl.txt").read().split()

Weights

All files are PyTorch dictionaries (torch.load(path, map_location="cpu")).

Ours (weights/mogloss/). Each .pt holds model (state dict) and config (d = 1024, 16 layers, 16 heads, one 207-d keypoint frame per token), and the encoder files also hold the input normalization mu / sd.

path	content
setting_{a,b}/pretraining/encoder_mmm_only.pt	stage 1, masked motion modeling
setting_{a,b}/pretraining/mmm_joint_full.pt	stage 2, joint motion and pseudo-gloss model (with the pseudo-gloss vocabulary)
setting_{a,b}/pretraining/encoder_mmm_joint.pt	encoder after both stages (our full model)
setting_{a,b}/pretraining/joint_only_full.pt, encoder_joint_only.pt	ablation, joint stage from random weights
setting_{a,b}/pretraining/encoder_no_pretraining.pt	ablation, random initialization
setting_a/translation/<encoder>/youtube_asl_pretrain.pt	translation model (encoder + T5-base) after YouTube-ASL
setting_a/translation/<encoder>/finetune_{how2sign,stem,openasl}.pt	fine-tuned per dataset (last checkpoint)
setting_b/translation/<encoder>/all_data.pt	one translation model on all data
setting_b/translation/data_size/{mmm_joint,mmm_only}_{50,25}pct.pt	translation stage on 50 % / 25 % of the data
islr/asl_citizen/{no_pretraining,mmm_only,mmm_joint}.pt	ASL Citizen classifiers (with the class list)
islr/wlasl{100,2000}/mmm_joint.pt	WLASL classifiers

<encoder> is one of mmm_joint, mmm_only, joint_only, no_pretraining.

Reproductions (weights/reproductions/), trained under the same protocol and data as our model.

path	content
uni-sign/setting_a_pose/	Uni-Sign pose-only: YouTube-ASL translation pretraining and per-dataset fine-tuning
uni-sign/setting_a_pose_rgb/	Uni-Sign pose + RGB: RGB stage 2 on YouTube-ASL and per-dataset fine-tuning
uni-sign/setting_b/all_data.pth	Uni-Sign on all data
uni-sign/islr/wlasl{100,2000}.pth	Uni-Sign WLASL classifiers (released WLASL recipe)
youtube_asl_baseline/	YouTube-ASL baseline (T5-v1.1-base): pretraining and per-dataset fine-tuning

The Uni-Sign files load with the official Uni-Sign code.

Citation

If you use any part of this repository, please cite

@article{hong2026mogloss,
  title   = {MoGloss: Joint Motion and Pseudo-Gloss Pretraining for American Sign Language Understanding},
  author  = {Hong, Rui and Ko{\v{s}}eck{\'a}, Jana},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}

and the source datasets listed below.

Sources and licenses

The keypoints, captions and labels are derived from YouTube-ASL, YouTube-SL-25, OpenASL, ASL STEM Wiki, How2Sign, ASL Citizen, ASL SignBank and WLASL. No video is redistributed.

This repository is released under the CC BY-NC 4.0 license (attribution, non-commercial use). Parts derived from a source dataset are also subject to the license and terms of that dataset. Please cite the source datasets when you use the corresponding part of this repository.